A tiny object detection dataset for high-altitude open-water search and rescue

Hui Xu, Chenfei Hou, Henglai Wei, Hao Wang, Yidong Jin, Xiulin Niu, Haoyang Zhang, Hui Zhang, Tianshi XuarXiv 2026 · arXiv

TL;DR

TinyWOD is a new specialized dataset for high-altitude open-water search and rescue (OW-SAR), featuring over 19,000 annotated instances of lifebuoys that span just a few pixels. By introducing a more granular "tiny object" metric ($AP_{tiny}$), the researchers provide a rigorous benchmark for detecting targets at altitudes of 1.5–2.0 km, where traditional "small object" detection methods often fail.

What Problem Does This Paper Address?

In open-water search and rescue, efficiency is king. Fixed-wing aircraft often fly at high altitudes (above 1.5 km) to cover vast areas quickly. However, this creates a massive computer vision challenge: at that height, a life-saving target like a lifebuoy might only occupy 3 to 15 pixels in a high-resolution image.

Existing maritime datasets mostly focus on low-altitude drone footage where targets are relatively large. Furthermore, standard evaluation metrics like the COCO $AP_{small}$ (which covers anything up to 32x32 pixels) are too broad to distinguish between a "small" object and a "tiny" one that is barely a speck on the horizon. TinyWOD addresses this gap by providing data that reflects the actual, difficult imaging conditions of high-altitude SAR—including varying water colors, surface ripples, and lighting.

Methods

The researchers faced a logistical hurdle: flying at 1.5 km is expensive and subject to strict airspace regulations. To overcome this, they used a pinhole camera model to simulate high-altitude imagery. By scaling down the physical size of the lifebuoys and flying at lower altitudes (up to 120m), they achieved the exact same "projected object size" in pixels as a full-sized lifebuoy viewed from 1.5 km.

The data acquisition setup used to simulate high-altitude imagery.

Key methodological contributions include:

  • Two-Stage Annotation: A rigorous process involving four independent annotators followed by two reviewers to ensure the accuracy of tiny bounding boxes.
  • $AP_{tiny}$ Metric: A new evaluation metric specifically for objects with an Equivalent Side Length (ESL) of less than 15 pixels. This provides a much clearer picture of how models perform on the hardest-to-detect targets.
  • Diverse Data Collection: Data was gathered across three regions in China, capturing green, yellow, and blue water bodies under both day and night conditions.
Geographical distribution of the data collection sites.

Datasets

The primary contribution is TinyWOD, which consists of:

  • 11,359 images (1920 × 1080 resolution).
  • 19,505 annotated instances of lifebuoys.
  • Average ESL of 12.7 pixels, with nearly 77% of instances falling below the 15-pixel "tiny" threshold.

The authors compared TinyWOD against six major overwater datasets, including SeaDronesSee, MOBDrone, and Airbus Ship. As shown in the comparison below, TinyWOD's targets are significantly smaller and occupy a much tinier fraction of the total image area than any existing benchmark.

Comparison of average object size (ESL) and area ratio across different datasets.

Key Insights and Ideas

  • The Failure of $AP_{small}$: In TinyWOD, 99.75% of objects qualify as "small" by COCO standards. This makes the standard $AP_{small}$ metric useless for ranking models, as it doesn't isolate the performance on the truly "tiny" 3-15 pixel targets that define high-altitude SAR.
  • Spatial Distribution Matters: Analysis showed that while objects were distributed across the frame, they tended to cluster toward the center. This insight is crucial for researchers looking to optimize search patterns or regional proposal networks.
Spatial and size distributions of objects within the TinyWOD dataset.
  • Simulation as a Proxy: The study demonstrates that mathematical scaling of targets is a viable way to build high-altitude datasets without the prohibitive costs of high-altitude flight, provided the pixel footprint is accurately maintained.

Key Results

The researchers benchmarked seven state-of-the-art detectors, including YOLO variants and Transformer-based models (DETRs).

  • Best Overall Performer: OWRT-DETR achieved the highest Average Precision (AP) of 0.512 and the best $AP_{tiny}$ of 0.440.
  • Precision vs. Recall: YOLO26x showed the highest precision (0.789), while H-DETR-R50 led in recall (0.669).
  • Localization Difficulty: The significantly lower scores for $AP_{75}$ (precision at a high overlap threshold) compared to $AP_{50}$ suggest that while models can find the tiny objects, accurately localizing their exact boundaries remains a major challenge.
Visual comparison of TinyWOD targets against other maritime datasets.

Limitations and Future Work

The authors acknowledge several areas for growth:

  • Single Category: The dataset currently only features lifebuoys. Future versions should include other SAR targets like persons in water or small vessels.
  • Simulation Gap: While the pixel footprint is accurate, scaled-down targets at low altitudes may not perfectly replicate the atmospheric haze or specific sensor noise encountered at 1.5 km.
  • Geographic Scope: The data is limited to one province in China; expanding to different oceanic environments would improve model robustness.

TinyWOD is available on Kaggle and Zenodo, providing a much-needed playground for engineers working on the "needle in a haystack" problem of maritime rescue.

Summary updated

Abstract

Large aerial platforms often operate above 1.5 km to improve the efficiency of open water search and rescue (OW-SAR). However, long imaging distances, wide fields of view, and limited sensor resolution reduce objects to a few pixels, increasing their susceptibility to environmental interference and making detection harder than for conventional small objects. To our knowledge, existing datasets focus on low-altitude scenarios and do not support automatic detection at 1.5-2.0 km. We therefore present TinyWOD, a tiny object detection dataset for high-altitude OW-SAR using lifebuoys as detection targets. It comprises 11,359 manually annotated images and 19,505 annotated instances spanning diverse environments, lighting conditions, water colors, and surface ripples. Compared with existing datasets, TinyWOD contains much smaller objects and a substantially higher proportion of tiny instances; its average equivalent side length (ESL) is 12.7 pixels, with 76.96% of in stances having ESLs below 15 pixels, reflecting actual high-altitude water-rescue imaging conditions. We define APtiny, a tiny object detection metric, and benchmark mainstream and state-of-the-art networks using this and conventional metrics. All evaluated networks effectively detect tiny waterborne objects, demonstrating TinyWOD's reusability and utility. We expect this dataset to support research on tiny object detection in high-altitude OW-SAR.
BibTeX
@misc{xu2026tiny,
  author = {Hui Xu and Chenfei Hou and Henglai Wei and Hao Wang and Yidong Jin and Xiulin Niu and Haoyang Zhang and Hui Zhang and Tianshi Xu},
  title = {{A tiny object detection dataset for high-altitude open-water search and rescue}},
  year = {2026},
  eprint = {2611.22825},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.22825},
}
Figure